Detection of vanishing point sequence with temporal fluctuation

A method for detecting the vanishing point from sequential images is proposed. In existing methods, a method for detecting the vanishing point from a single image had been discussed. However, since sequential images are frequently used in environment recognition from images, it is important that the temporal fluctuation characteristics caused by camera vibration are considered. In this paper, we introduce such a temporal characteristic, and one vanishing point sequence is stably estimated based on both the temporal characteristic and the information on lines constructed from feature points as observed data. In our approach, the expectation‐maximization (EM) algorithm is used hierarchically, and both maximum a posteriori (MAP) estimation of the vanishing point sequence and line clustering for feature points are simultaneously performed. The saddle point approximation is used for convenience in calculating the expectation with respect to the posterior in the EM algorithm. Experiments in which autoregressive (AR) processing was applied to the temporal characteristics and 400 images were used confirmed that the error was effectively reduced in comparison with the conventional method for images with lower reliability of the observed data.